Qualcomm Technologies, Inc. has announced its latest data center AI inference optimization solutions, unveiling the Qualcomm® AI200 and AI250 chips built into accelerator cards and racks. Based on Qualcomm’s industry-leading NPU technology, these solutions deliver rack-scale performance with high memory capacity, enabling efficient generation-style AI inference at an exceptional watts-per-performance ratio. This represents a significant leap forward toward building scalable, efficient, and flexible generative AI infrastructure across industries.
The Qualcomm AI200 is a rack-level AI inference-focused solution designed to minimize total cost of ownership (TCO) while optimizing performance for a wide range of AI workloads, including large language models (LLMs) and multi-modal models (LMMs). It supports up to 768GB of LPDDR memory per card, providing both high memory capacity and cost efficiency, along with scalability and flexibility critical for large AI inference environments.
Meanwhile, the Qualcomm AI250 employs a groundbreaking memory architecture based on near-memory computing technology, significantly boosting memory bandwidth by over ten times and drastically reducing power consumption. This advancement enhances the efficiency and performance of AI inference workloads across multiple generations. The solution also supports a disaggregated inference architecture, allowing more effective hardware resource utilization and meeting diverse performance and cost requirements.
Both AI200 and AI250 solutions feature direct liquid cooling for improved thermal efficiency, support PCIe and Ethernet for scalability, and integrate confidential computing for secure AI workload processing. The rack-level power consumption for these solutions is approximately 160kW.
Dongura Maladi, Senior Vice President and General Manager of Qualcomm Technologies' Data Center and Edge Solutions, stated, “With AI200 and AI250, we are redefining rack-scale AI inference capabilities. These innovative infrastructure solutions enable deployment of generative AI with outstanding TCO and provide the scalability and security data centers demand.” He added, “Through Qualcomm’s comprehensive software stack and open ecosystem support, developers and enterprises can easily integrate, manage, and scale pre-trained AI models with our optimized inference solutions. Complete compatibility with major AI frameworks and one-click model deployment deliver a seamless experience, fostering rapid innovation.”
Qualcomm’s hyperscaler-grade AI software stack, optimized for AI inference, covers application to system software layers end-to-end. It supports major machine learning (ML) frameworks, inference engines, generative AI frameworks, and disaggregated inference optimization techniques for LLMs and LMMs. Developers can deploy Hugging Face models via Qualcomm’s Efficient Transformers Library and Qualcomm AI Inference Suite with one click, simplifying model onboarding. Additionally, Qualcomm’s software provides ready-to-use AI applications, agents, tools, libraries, APIs, and services to facilitate real-world AI applications.
The AI200 and AI250 are scheduled for commercial release in 2026 and 2027, respectively. Qualcomm continues to focus on delivering industry-leading AI inference performance, energy efficiency, and optimal TCO through its annual updated data center roadmap. More detailed information is available on Qualcomm’s official website.
